Professional Documents
Culture Documents
These notes follow closely the development of inner constraint equations by Dr Willie Tan,
Department of Building, School of Design and Environment, National University of Singapore (Tan
2005).
v + Bx = f (1)
Nx = t (2)
where
N = BT WB and t = BT Wf (3)
If N is non-singular, i.e., N ≠ 0 and N −1 exists: there is a solution for the unknowns x given by
matrix inversion
x = N − 1t (4)
If N is singular, i.e., N = 0 and N −1 does not exist: there is no solution by conventional means
The aim of these notes is to explain how we might obtain least squares solutions where the normal
coefficient matrix is singular. Such solutions, using INNER CONSTRAINTS are known as FREE
NET ADJUSTMENTS.
In least squares adjustments of survey networks, singular sets of normal equations (singular normal
coefficient matrices N) are rank deficient sets of equations and arise because design matrices B are
rank deficient. This will invariably be due to datum defects, which could be: coordinate origin not
defined (no fixed points), the network orientation not defined (no lines with fixed directions in
space), no scale (no measured distances) or no height datum defined (no points with fixed heights).
The datum defects are directly connected to the rank deficiency of the design matrix B.
A way to overcome this datum deficiency and find a solution is to impose constraints, and these
constraints take the form of a set of CONSTRAINT EQUATIONS
Cx = g (5)
Combining the constraint equations (5) with the observation equations (1) and enforcing the least
squares condition
ϕ = vT Wv − 2k T ( Cx − g ) ⇒ minimum (6)
where W is the n × n weight matrix and k is the c × 1 vector of Lagrange multipliers, gives rise to
the set of equations
⎡ − N CT ⎤ ⎡ x ⎤ ⎡ − t ⎤
⎢ ⎥ =
0 ⎥⎦ ⎢⎣ k ⎥⎦ ⎢⎣ g ⎥⎦
(7)
⎢⎣ C
Equations (7) can be solved provided that C consists of linearly independent rows that overcome
the datum problems. For a particular network with the maximum number of datum defects (e.g., a
height network with no fixed points, a 2-D network of distances and directions with no fixed points
or fixed orientation) there are many different constraints that can be imposed to enable a solution
for x, and amongst this family of possible constraints there are MINIMAL CONSTRAINT sub-sets,
i.e., sub-sets containing the minimum number of constraints required for a solution; the minimum
number being equal to the rank deficiency. Within this group of minimal constraint sub-sets there
will be a single set called INNER CONSTRAINTS. This set of inner constraints (inner constraint
equations) has the unique property that Cx = 0 and a solution of a network using inner constraints
yields x such that the norm of x, denoted as x , is a minimum and the trace of Q xx denoted as
tr ( Q xx ) is also minimum. The trace of a (square) matrix is the sum of the elements of the leading
diagonal, and Q xx is the (square) cofactor matrix containing estimates of the variances s x2k of the
adjusted quantities (the elements of x). Hence, a minimum trace solution has a set
{s 2
x1 s x22 }
s x2u whose sum of the u elements is the smallest, i.e., a minimum variance solution.
The norm of a column (or row) vector is the square root of the sum of squares of the elements, i.e.,
the "length" of the vector. A minimum norm solution would have a vector x having the smallest
length.
Adjustments of survey networks where no datum is defined and INNER CONSTRAINTS are
employed, are known as FREE NET ADJUSTMENTS, and those employing MINIMAL
CONSTRAINTS are used are known as MINIMALLY CONSTRAINED ADJUSTMENTS.
In free net adjustments (where every point in the network is free to move) the inner constraints have
the unique property that Cx = 0 , i.e., certain linear functions of the adjusted quantities x are equal
to zero. These adjusted quantities can be heights in level networks, or coordinates and theodolite
orientation constants in conventional 2-D and 3-D survey networks, or just coordinates in
photogrammetric and GPS networks. In conventional adjustments (where there are enough fixed
points to properly define the datum and orientation) or minimally constrained adjustments, or
conventional adjustments with additional constraints, the constraint equations are conditions
relating to actual points in the network. In free net adjustments the constraints are applied to, or
relate to, a single "fictitious" point; the centroid of the network. The key here is to remember that
the most convenient way to represent position is via height (in level networks) or coordinates in 2-
D and 3-D networks, hence most network adjustments are actually variation of coordinate
adjustments; height being a 1-D system. So, coordinates (approximate or fixed) are required for all
points in the network – even if the coordinate datum or height datum is not fixed, or the orientation
is not properly known. Once these coordinates are assigned then the centroid is defined, since it is
just the average of the coordinates of all of the points. We can now decide on certain constraints on
the centroid and between the centroid and all the points in the network that will overcome datum
defects in 1-D, 2-D and 3-D survey networks. These three constraints are:
• the centroid remains unchanged;
• the average bearing from the centroid to all points remains unchanged;
• the average distance from the centroid to all points remains unchanged.
These constraints give rise to positional constraints, rotational constraints and a scale constraint.
In a free network adjustment of heights (a 1-D system) there is one datum defect and so one
positional constraint is required and this is derived from the condition that the centroid remains
unchanged. In a free net adjustment of a 2-D network (measured directions and distances) there are
three datum defects, so two positional constraints and one rotational constraint are required. These
are derived from the condition that the centroid remains unchanged and the average bearing from
the centroid to all points remains unchanged. For 3-D networks (measured horizontal directions,
zenith angles and distances) there are six datum defects, requiring three positional and three
rotational constraints. Again, these are derived from the condition that the centroid remains
unchanged and that the average bearing (in space) from the centroid to all points remains
unchanged. For 2-D and 3-D networks without measured distances, there is an additional scale
constraint required which can be derived from the condition that the average distance from the
centroid to all points remains unchanged.
The positional, rotational and scale (inner) constraints for a 3-D survey network are derived as
follows.
POSITIONAL CONSTRAINT
Let ( X k , Yk , Z k ) be the coordinates of an arbitrary point Pk and there are m points in the network.
X + X2 + X3 + + Xm ∑ Xk ∑Yk ∑Z k
XG = 1 = k =1
; YG = k =1
; ZG = k =1
(8)
m m m m
If we impose the condition that G does not move, then there can be no change to the centroid
coordinates during an iterative least squares solution, hence δ X G = δ YG = δ Z G = 0 . Now since
∂X G ∂X ∂X G
δ XG = δ X1 + G δ X 2 + + δ Xm
∂X 1 ∂X 2 ∂X m
1 1 1
= δ X1 + δ X 2 + + δ Xm
m m m
m
1
= ∑δ X k
m k =1
1 m 1 m
And similarly, δ YG = ∑
m k =1
δ Yk and δ Z G = ∑ δ Z k . Remembering that δ X G = δ YG = δ Z G = 0 , the
m k =1
positional constraints are
m m m
∑δ X k = 0 ;
k =1
∑δ Yk = 0 ;
k =1
∑δ Z
k =1
k =0 (9)
ROTATIONAL CONSTRAINTS
The three rotational constraints can be derived by imposing the condition that the average bearing
from the centroid G to all points remains unchanged.
Figure 1 shows a line in 3-D space from G to an arbitrary point Pk ( X k , Yk , Z k ) . The line GPk has
length sk and projections rk = GPk′ , qk = GPk′′ and pk = GPk′′′ on to the X-Y, Y-Z and X-Z planes
respectively.
Consider the projection rk = GPk′ on the X-Y plane that has bearing θ k . Plane trigonometry gives
⎛ Xk − XG ⎞
θ k = tan −1 ⎜ ⎟ = θ k ( X k , Yk , X G , YG ) (10)
⎝ Yk − YG ⎠
|| Z
P'''
k P''
k
ψ qk
p
k
φ
Pk X k ,Yk , Zk
G sk
θ
rk
|| Y
|| X
| P'k
Figure 1. A line in space and its projection on the X-Y, Y-Z and X-Z planes.
Noting that X G , YG in equation (10) are constant (since the centroid cannot move), the Total
∂θ k ∂θ
δθ k δ X k + k δ Yk (11)
∂X k ∂Yk
∂θ k cos θ k Yk − YG ∂θ k sin θ k X −X
= = ; =− =− k 2 G
∂X k rk rk2 ∂Yk rk rk
Hence
δθ k
1
rk2
{(Yk − YG ) δ X k − ( X k − X G ) δ Yk } (12)
If the average bearing from G to all m points in the network is to remain unchanged then
m
∑δθ
k =1
k =0 (13)
m
Let ∑δθ
k =1
k be the vector s = [δθ1 + δθ 2 + + δθ m ] and s contains a single scalar quantity. In
equation (13) s = [0] = 01×1 . Now let s be the product of two vectors
s = rT d (14)
⎡1⎤
⎢ r2 ⎥
⎢1 ⎥ ⎡ (Y1 − YG ) δ X 1 − ( X 1 − X G ) δ Y1 ⎤
⎢1⎥ ⎢
(Y2 − YG ) δ X 2 − ( X 2 − X G ) δ Y2 ⎥⎥
where r = ⎢ r22 ⎥ and d= ⎢
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢1⎥ ⎣(Ym − YG ) δ X m − ( X m − X G ) δ Ym ⎦ m×1
⎢ 2⎥
⎢⎣ rm ⎥⎦ m×1
rs = rr T d
rr T is a matrix of order m × m of full rank, since the elements of r are independent and the inverse
(rr ) rs = ( rr T ) (rr ) d = Id = d
T −1 −1 T
(15)
When s on the left-hand-side of equation (15) is zero (enforcing the condition of equation (13)) then
L.H.S. = 0m×1 and every element of d m×1 on the R.H.S. will be equal to zero, giving
∑{(Y
k =1
k − YG ) δ X k − ( X k − X G ) δ Yk } = 0 (16)
∑Ykδ X k − ∑YGδ X k − ∑ X kδ Yk + ∑ X Gδ Yk = 0
k =1 k =1 k =1 k =1
m m m
∑ (Y δ X
k =1
k k − X kδ Yk ) − YG ∑ δ X k + X G ∑ δ Yk = 0
k =1 k =1
m m
But from equation (9), the positional constraints, ∑δ X
k =1
k = ∑δ Y
k =1
k = 0 hence the rotational
∑ (Y δ X
k =1
k k − X kδ Yk ) = 0 (17)
[Equations (16) and (17) are identical to Tan 2005 (eq'ns (12) and (13), p. 93)]
Now, referring to Figure 1, the projection qk = GPk′′ on the Y-Z plane makes an angle φ with the Z-
⎛ Yk − YG ⎞
φk = tan −1 ⎜ ⎟ = φk (Yk , Z k , YG , Z G ) (18)
⎝ Z k − ZG ⎠
∂φk ∂φ
δφk δ Yk + k δ Z k
∂Yk ∂Z k
δφk
1
qk2
{( Z k − ZG ) δ Yk − (Yk − YG ) δ Z k }
If the average bearing from G (the centroid) to all m points in the network is to remain unchanged
then the rotational constraint in the Y-Z plane is
m
∑δφ
k =1
k =0 (19)
In a similar manner as before, the rotational constraint in the Y-Z plane can be expressed as
∑(Z δY
k =1
k k − Ykδ Z k ) = 0 (20)
Again, referring to Figure 1, the projection pk = GPk′′′ on the X-Z plane makes an angle ψ with the
⎛ Xk − XG ⎞
ψ k = tan −1 ⎜ ⎟ = ψ k ( X k , Z k , X G , ZG ) (21)
⎝ Z k − ZG ⎠
∑δψ
k =1
k =0 (22)
which becomes
∑(Z δ X
k =1
k k − X kδ Z k ) = 0 (23)
Equations (13), (19) and (22) are rotational constraints expressed in angular quantities. Equations
(17), (20) and (23) are the same rotational constraints expressed in linear quantities (coordinates
and coordinate corrections)
SCALE CONSTRAINT
sk = ( Xk − XG ) + (Yk − YG ) + ( Z k − Z G ) = sk ( X k , Yk , Z k , X G , YG , Z G )
2 2 2
(24)
Noting that X G , YG , Z G in equation (24) are constant (since the centroid cannot move), the Total
∂sk ∂s ∂s
δ sk δ X k + k δ Yk + k δ Z k (25)
∂X k ∂Yk ∂Z k
δ sk
1
sk
{( X k − X G ) δ X k + (Yk − YG ) δ Yk + ( Z k − ZG ) δ Z k } (26)
If the average distance from G (the centroid) to all m points in the network is to remain unchanged
then
m
∑δ s
k =1
k =0 (27)
In the same manner as for the development of rotational constraints, matrix manipulations can be
used to turn the sum of small changes in distances to a sum of small changes in coordinates and
equations (26) and (27) imply
m
∑{( X
k =1
k − X G ) δ X k + (Yk − YG ) δ Yk + ( Z k − Z G ) δ Z k } = 0 (28)
∑( X δ X
k =1
k k + Yk δ Yk + Z k δ Z k ) − X G ∑ δ X k − YG ∑ δ Yk − Z G ∑ δ Z k = 0
k =1 k =1 k =1
m m m
But from equation (9), the positional constraints, ∑δ X k =
k =1
∑ δ Yk = ∑ δ Z k = 0 hence the scale
k =1 k =1
∑( X δ X
k =1
k k + Yk δ Yk + Z k δ Z k ) = 0 (29)
[Equations (28) and (29) are identical to Tan 2005 (eq'ns (20) and (21), p. 93)]
SUMMARY OF CONSTRAINTS
Positional constraints
m m m
∑δ X k = 0 ;
k =1
∑δ Yk = 0 ;
k =1
∑δ Z
k =1
k =0
Rotational constraints
m
∑ (Y δ X
k =1
k k − X k δ Yk ) = 0
m
∑(Z δY
k =1
k k − Yk δ Z k ) = 0
m
∑(Z δ X
k =1
k k − X kδ Zk ) = 0
Scale constraint
m
∑( X δ X
k =1
k k + Yk δ Yk + Z k δ Z k ) = 0
Equation (7) is the matrix equation for a constrained least squares solution of a survey network,
where C is the coefficient matrix of the constraint equations Cx = g . In the case of INNER
CONSTRAINTS, Cx = 0 and equation (7) becomes
⎡ − N CT ⎤ ⎡ x ⎤ ⎡ − t ⎤
⎢ ⎥ ⎢k ⎥ = ⎢ 0⎥ (30)
⎣⎢ C 0 ⎥⎦ ⎣ ⎦ ⎣ ⎦
In the case of a 3-D network, with no measured distances (no scale) and no fixed points there will
be seven datum defects requiring seven constraint equations to overcome the defects in a free net
adjustment. The seven constraints combine (i) three positional constraints, (ii) three rotational
constraints and (iii) one scale constraint. If this network solution has the coordinate corrections in x
in the following sequence
xT = [δ X 1 δ Y1 δ Z1 δ X 2 δ Y2 δ Z 2 δ X m δ Ym δ Z m ]
⎡ δ X1 ⎤
⎢ δY ⎥
⎡1 0 0 1 0 0 1 0 0 ⎤⎢ 1 ⎥
⎢0 ⎢δZ ⎥
1 0 0 1 0 0 1 0 ⎥⎢ 1 ⎥
⎢ ⎥ δX
⎢0 0 1 0 0 1 0 0 1 ⎥⎢ 2⎥
⎢ ⎥ ⎢ δY ⎥
⎢ Y1 − X1 0 Y2 −X2 0 Ym −Xm 0 ⎥⎢ 2 ⎥ = 0 (31)
δZ
⎢0 Z1 −Y1 0 Z2 −Y2 0 Zm −Ym ⎥ ⎢ 2 ⎥
⎢ ⎥⎢ ⎥
⎢ Z1 0 − X1 Z2 0 −X2 Zm 0 −Xm ⎥ ⎢ ⎥
⎢ X1 ⎢ δ X ⎥
⎣ Y1 Z1 X2 Y2 Z2 Xm Ym Z m ⎥⎦ m
⎢ δY ⎥
⎢ m⎥
⎣⎢ δ Z m ⎦⎥
In subsequent notes on this topic, several examples of free net adjustments will be shown.
REFERENCES
Tan, W., 2005, 'Inner constraints for 3-D survey networks', Spatial Science, Vol. 50, No. 1, June
2005, pp. 91-94.